Teeth2Point: A Two-Stage Dental CBCT ROI-to-Point Segmentation Framework
本文提出Teeth2Point框架，通过两阶段方法解决牙齿CBCT图像中缺失或错位牙齿的准确标注问题，先用卷积模型定位ROI，再用转换器处理点令牌。
0 citationsRead paper
Selected work
Representative Papers
Event-Driven On-Sensor Locomotion Mode Recognition Using a Shank-Mounted IMU with Embedded Machine Learning for Exoskeleton Control
This work proposes an energy-efficient approach to lower-limb exoskeleton control by embedding a lightweight decision tree model directly within the inertial measurement unit (IMU)—specifically the ST LSM6DSV16X on the shank—leveraging its built-in Machine Learning Core (MLC) for event-driven, on-chip activity recognition. Instead of continuously transmitting raw inertial data to the main microcontroller for motion classification, the IMU processes sensor data locally and sends only the classification result to the exoskeleton controller. This is the first implementation of interrupt-driven motion mode recognition deployed entirely on the IMU side, eliminating the need for custom machine learning code on the host processor. The method significantly reduces system power consumption, communication overhead, and latency while maintaining high recognition robustness across three key locomotion modes: standing, level-ground walking, and stair climbing, thereby extending battery life and enabling low-latency assistive control.
0 citationsRead paper
Recent publications
Latest Papers
Teeth2Point: A Two-Stage Dental CBCT ROI-to-Point Segmentation Framework
本文提出Teeth2Point框架，通过两阶段方法解决牙齿CBCT图像中缺失或错位牙齿的准确标注问题，先用卷积模型定位ROI，再用转换器处理点令牌。
0 citationsRead paper
Event-Driven On-Sensor Locomotion Mode Recognition Using a Shank-Mounted IMU with Embedded Machine Learning for Exoskeleton Control
This work proposes an energy-efficient approach to lower-limb exoskeleton control by embedding a lightweight decision tree model directly within the inertial measurement unit (IMU)—specifically the ST LSM6DSV16X on the shank—leveraging its built-in Machine Learning Core (MLC) for event-driven, on-chip activity recognition. Instead of continuously transmitting raw inertial data to the main microcontroller for motion classification, the IMU processes sensor data locally and sends only the classification result to the exoskeleton controller. This is the first implementation of interrupt-driven motion mode recognition deployed entirely on the IMU side, eliminating the need for custom machine learning code on the host processor. The method significantly reduces system power consumption, communication overhead, and latency while maintaining high recognition robustness across three key locomotion modes: standing, level-ground walking, and stair climbing, thereby extending battery life and enabling low-latency assistive control.
0 citationsRead paper